AI Red Teaming: Identity, Access, and Attack Paths Still Matter

This article is an abridged interview discussing how common security failures (weak identity, excessive privileges, exposed tokens, and insufficient separation of duties) become more consequential when connected to AI systems; it emphasizes prompt injection as a novel testing concern, the nondeterministic nature of model behavior, and practical mitigations such as inventorying non-human identities, reviewing effective access, mapping attack paths, and reducing default trust.